Artificial intelligence for opportunistic screening of osteoporosis across multiple imaging modalities: a systematic review Authors: Xiang Chen, Wanling Wu, Jianying Shen, Yuyan Chen Published: 2026 ABSTRACT BackgroundOsteoporosis often remains undetected until fracture. Dual-energy X-ray absorptiometry (DXA) screening is limited by poor accessibility. Opportunistic screening of routine medical images, powered by artificial intelligence (AI), may enable automated, large-scale bone health assessment.MethodsThis systematic review followed PRISMA 2020 guidelines. We searched PubMed, Cochrane Library, and Web of Science up to February 2026 for studies developing AI models for opportunistic osteoporosis screening. Two reviewers independently screened, extracted data, and assessed risk of bias using phase classification and PROBAST. Outcomes included osteoporosis, osteopenia, and fracture risk prediction, evaluated mainly by area under the curve (AUC). Principal component analysis (PCA) was used to explore sources of heterogeneity.ResultsOf 57 included studies (2019–2026), most were retrospective and single-center. Computed tomography (CT) was the most common modality (38 studies), and the spine was the most frequent ROI. Deep learning (29 studies) has largely replaced traditional machine learning, and foundation models have recently emerged. Model performance was generally high (AUC range 0.630–1.000 across tasks), but heterogeneity was substantial. PCA identified sample size, number of centers, and single-center design as major drivers of heterogeneity, smaller and single-center studies tended to report higher AUCs, suggesting possible overestimation. PROBAST assessment revealed that the absence of external validation was the leading source of bias, only 36% of studies conducted external validation, with additional concerns in participant selection and analysis domains.ConclusionAI-based opportunistic osteoporosis screening has progressed rapidly with encouraging performance. However, the evidence base remains dominated by retrospective single-center studies with limited external validation, which hinders clinical translation. Future work should prioritize multicenter collaboration, prospective validation, and standardized reporting.Systematic review registrationhttps://www.crd.york.ac.uk/PROSPERO/view/CRD420261451716, identifier PROSPERO (CRD420261451716). Introduction Osteoporosis is one of the most prevalent skeletal diseases worldwide (1, 2). It is characterized by low bone mass and deterioration of bone microarchitecture, which increase bone fragility and susceptibility to fracture. In the United States, osteoporosis affects ~26% of women aged ≥65 years and more than 50% of those aged ≥85 years, with an even higher prevalence among postmenopausal white women (3). Globally, nearly 1/3 of women and 1/5 of men older than 50 years will experience an osteoporotic fracture during their lifetime (4). Fragility fractures, particularly of the hip and vertebrae, are the main clinical consequence and are associated with substantial morbidity, increased mortality, and considerable healthcare costs (5–7). Osteoporosis often remains undetected until a fracture occurs (8). Although diagnostic criteria are well-established, underdiagnosis and undertreatment remain common in clinical practice. At present, diagnosis mainly relies on bone mineral density (BMD) measured by dual-energy X-ray absorptiometry (DXA), with a T-score ≤ − 2.5 indicating osteoporosis (9). However, limited access and imperfect sensitivity reduce its value as a population screening tool (10). A meta-analysis of 19 studies showed that quantitative computed tomography (QCT) identified substantially more osteoporosis cases than DXA (11). Screening rates in routine care also remain low. Among patients with upper-extremity fragility fractures, only 51% underwent DXA, 25% were referred for endocrine evaluation, and 20% initiated pharmacological treatment, with markedly lower DXA utilization in men than women (18% vs. 55%) (12). Screening gaps are also evident in younger individuals. In a fracture liaison service cohort of 1,248 patients, 61.9% were younger than 65 years, yet only 34% met the International Society for Clinical Densitometry (ISCD) criteria for BMD testing (13). These findings suggest that current screening strategies may fail to capture a substantial proportion of at-risk individuals. To address these limitations, opportunistic screening has gained increasing attention (13, 14). This strategy uses imaging obtained for other clinical purposes to evaluate bone health during routine care, without additional radiation exposure or extra patient visits. Its value is supported by the widespread use of computed tomography (CT), especially in older adults, who frequently undergo chest, abdominal, or pelvic CT scans for unrelated indications. Reanalyzing these existing scans can help identify patients at high risk of osteoporosis with minimal additional burden. CT remains the most extensively studied modality for opportunistic screening. Since a landmark study in 2013 established its feasibility (15), subsequent research has confirmed its diagnostic performance (16–19). More recently, advances in image analysis have extended opportunistic screening to conventional radiographs and magnetic resonance imaging (MRI), thereby further broadening its clinical applicability (20). In parallel, artificial intelligence (AI), particularly deep learning, has rapidly advanced this field (21, 22). Although clinical adoption remains limited, AI offers important advantages by automating key steps in the screening workflow, including region-of-interest (ROI) selection, image segmentation, BMD estimation, and risk prediction. This may improve efficiency, consistency, and reproducibility. More recently, task-agnostic foundation models, which have become standard in natural language processing and computer vision, have begun to attract attention in medical image analysis. Transfer learning from such models has proven effective for segmentation and classification tasks (23), and models such as CheXagent and DINOv2 are now being explored for chest radiograph interpretation. In osteoporosis diagnosis, applications remain scarce but are emerging. For instance, a recent study compared natural-domain and medical-domain foundation models for opportunistic osteoporosis screening and found that, after appropriate fine-tuning, the medical-domain models did not outperform their natural-domain counterparts (24). This finding challenges the common assumption that domain-specific pretraining is always necessary, and suggests that fine-tuned natural-domain foundation models may suffice for certain medical imaging tasks, potentially lowering data and computation barriers. Despite this promise, the role of foundation models in opportunistic osteoporosis screening has not been systematically reviewed. Previous reviews have highlighted the promise of AI-based CT screening. One review, which searched the literature up to December 2023, identified 14 studies and reported high accuracy for BMD prediction and osteoporosis classification. However, its relatively early search end date means that a substantial number of more recent publications were not captured. A more recent review included 55 studies, but its scope was broader and not limited to opportunistic screening (25, 26). Importantly, both reviews focused exclusively on CT, whereas our review additionally synthesizes evidence from radiography and MRI, including DWI and ADC MRI, which have shown promising predictive performance in the past 2 years (27, 28). This review examines the use of AI for opportunistic osteoporosis screening across multiple imaging modalities. We summarize the characteristics of existing studies, evaluate their methodological quality, explore major sources of heterogeneity, and discuss barriers to clinical implementation and future research priorities. By providing a comprehensive overview of the current evidence, this review aims to clarify the status of the field and highlight the clinical potential of AI-driven opportunistic screening. Materials and methods This systematic review was conducted in accordance with the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA 2020) guidelines. The complete PRISMA checklist is provided in Supplementary Table 1. Search strategy A comprehensive literature search was performed in PubMed, Cochrane Library, and Web of Science from inception to February 28, 2026. The search strategy was developed by the research team and combined Medical Subject Headings (MeSH) terms with free-text keywords. The full search strategy is detailed in Supplementary Table 2. Eligibility criteria (PICOS framework) The eligibility criteria were established according to the PICOS (Population, Index test, Comparator, Outcomes, and Study design) framework. The population comprised individuals undergoing routine medical imaging examinations for whom opportunistic osteoporosis screening was applicable. The index test consisted of artificial intelligence-based algorithms applied to medical images. The comparator included DXA or QCT as the reference standard. The primary outcome of interest was the area under the receiver operating characteristic curve (AUC), which served as the quantitative metric for diagnostic performance. Eligible studies included retrospective or prospective studies reporting the development of AI models for opportunistic osteoporosis screening. Study selection Title and abstract screening were conducted independently by two reviewers (Xiang Chen and Wanling Wu) according to predefined exclusion criteria. Records were excluded if they were reviews, editorials, conference abstracts, studies unrelated to model development, studies not focused on osteoporosis screening, studies using non-standard imaging modalities, or non-English publications. Inter-reviewer agreement was assessed using Cohen’s Kappa, calculated with the “kappa2” function from the R package “irr” (v0.84). The agreement for this phase was good (Kappa = 0.81, 95% confidence interval (CI): 0.73–0.89). For full-text screening, the complete texts of 134 articles were retrieved and independently assessed by the same two reviewers (Xiang Chen and Wanling Wu). Further exclusions were made for the following reasons: non-opportunistic screening, absence of model development, missing model performance metrics, unavailability of the full text, or inconsistent reporting of participant numbers. Inter-reviewer agreement was again evaluated using Cohen’s Kappa, yielding a value of 0.86 (95% CI: 0.77–0.95), indicating very good agreement. All disagreements were resolved through discussion or, when necessary, by consultation with a third reviewer (Yuyan Chen). The remaining studies were included in the final analysis. Data extraction A standardized data extraction form was developed prior to data collection to capture information on data source, sample size, study population, study design, imaging modality, research question, image segmentation approach, clinical and imaging features, validation method, reference standard, model type, etc. Two reviewers (Xiang Chen and Wanling Wu) independently extracted data from each included study. Any disagreements were resolved through discussion, with a third reviewer (Yuyan Chen) consulted when consensus could not be reached. The primary outcomes of interest were the prediction of osteoporosis, osteopenia, or fracture risk. From each included study, the area under the curve (AUC) was extracted as the primary metric of model performance. When multiple models or AUC values were reported in a single study, only one AUC was selected according to the following pre-specified hierarchical criteria (applied in order): (1) The AUC of the primary or main model explicitly designated by the authors (e.g., described as the “primary model,” “final model,” or “best-performing model”); (2) The AUC from external validation (independent test cohort), prioritized over internal validation or development set results. If both internal and external validation results were reported, the external validation AUC was extracted even if it was lower than the internal AUC; (3) For studies reporting multiple external validations, the AUC from the validation cohort most similar to our target setting, or the pooled result across external sets if presented by the authors; (4) As a last resort, when none of the above criteria could be applied, the highest reported AUC. This hierarchical approach was designed to balance methodological rigor with the need for consistent data extraction across all 57 included studies. We acknowledge that selecting the highest AUC in certain cases may still risk some overestimation of performance. This limitation is discussed further in the Limitations section. Quality assessment The methodological quality of all included studies was independently assessed by two reviewers (Xiang Chen and Wanling Wu), with cross-checking to ensure consistency. Inter-reviewer agreement was quantified using Cohen’s Kappa for binary classifications, computed with the “kappa2” function from the R package “irr” (v0.84). All disagreements were resolved through discussion or, when necessary, by consultation with a third reviewer (Yuyan Chen). Model quality assessment Model quality was evaluated using the Phase classification criteria, which consider sample size (<100 or ≥100), study design (retrospective or prospective), validation approach (internal or independent/external), and development phase (pre-market or post-market). According to this framework, imaging-based studies are typically classified as Phase II or III. Inter-reviewer agreement was 98.25%, with a Kappa of 0.66. Risk of bias assessment The risk of bias and applicability of the included studies were assessed using the Prediction model Risk Of Bias Assessment Tool (PROBAST). This tool evaluates four domains: participants, predictors, outcome, and analysis, and provides an overall judgment of risk of bias. Each domain contains 2, 3, 6, and 9 signaling questions, respectively. Responses are categorized as “yes/probably yes,” “no/probably no,” or “no information.” A domain was rated as high risk of bias if at least one signaling question was answered “no” or “probably no.” The overall risk of bias was considered low only if all four domains were judged to be at low risk. Inter-reviewer agreement for the overall ROB judgment was good, with a Cohen’s Kappa of 0.75 (95% CI: 0.58–0.92, observed agreement = 87.72%). Principal component analysis (PCA) To investigate potential sources of heterogeneity, study-level principal component analysis (PCA) was performed separately for studies predicting osteoporosis and osteopenia. Thirteen study characteristics were included: three numerical variables (AUC, number of participants, and number of centers) and ten categorical variables (prediction outcome, reference standard, external validation status, multicenter status, country, study design, gender, region of interest category, imaging modality, and feature extraction method). Numerical variables were standardized, and categorical variables were one-hot encoded. Each study was treated as an independent observation. PCA was conducted using the “prcomp” function in the R package stats (version 3.6.2), and the results were visualized with the “fviz_pca_biplot” function in the R package “factoextra” (version 2.0.0). Results Literature search and study selection A comprehensive literature search was conducted in PubMed, Cochrane, and Web of Science (last updated February 28, 2026). The detailed search strategy and key terms were described in the Methods section and Supplementary Table 2. In addition, the reference lists of relevant articles were manually screened to identify potentially eligible studies. A total of 383 records were initially identified (154 from PubMed, 8 from Cochrane Library, and 221 from Web of Science). After removal of duplicates, 265 unique records remained and were subjected to title and abstract screening. Of these, 133 records were excluded for the following reasons: conference abstracts (n = 29), reviews or editorials (n = 45), non-English publications (n = 1), non-routine imaging modalities (n = 1), absence of artificial intelligence model development (n = 35), and irrelevance to osteoporosis screening (n = 22). The remaining 132 articles, supplemented by 2 additional studies identified during the peer review process, brought the total to 134 articles that proceeded to full-text assessment. During full-text review, 77 articles were excluded due to: non-opportunistic screening design (n = 21), inconsistency between reported participant numbers and model development cohorts (n = 1) (29), focus on applying existing AI tools rather than developing new models (n = 38), lack of clear performance metrics (n = 16), and unavailability of the full text (n = 1). Finally, 57 studies met the inclusion criteria and were included in the systematic review. Study selection was performed independently by two reviewers, and discrepancies were resolved through discussion and consensus, resulting in a high level of inter-reviewer agreement. The study selection process is illustrated in the PRISMA flow diagram (Figure 1). PRISMA flowchart of included studies. Study and population characteristics: rapid growth with substantial heterogeneity Study design characteristics A total of 57 studies were included. Detailed information is provided in Table 1. Publication years ranged from 2019 to 2026. The majority of studies (n = 45, ~78%) were published from 2023 onward (30–72), with a particularly high concentration in 2025 and early 2026 (n = 19), indicating that AI-based opportunistic osteoporosis screening has entered a phase of rapid expansion. Geographically, the studies were dominated by mainland China, which contributed approximately 47% of all studies (n = 26) (34–36, 39–42, 44, 45, 47, 48, 51, 52, 55, 57, 59–62, 64, 67, 69, 71, 73–75). Other contributing countries included the United States (n = 7) (46, 65, 68, 76–79), South Korea (n = 6) (38, 50, 53, 54, 80, 81), and Italy (n = 3) (31, 43, 82), with smaller numbers of studies originating from Taiwan (33, 56, 58, 66, 83), Finland (84), Germany (85), Japan (32, 63), Spain (30), Switzerland (49), Turkey (72, 86), as well as two multi-region studies (37, 70). This geographic imbalance may reflect regional differences in imaging utilization and varying levels of institutional and policy support for AI adoption. Table 1Characteristics of study design.IDFirst authorPublication yearParticipantTrainingValidationTestingCross-validationExternal validationMulti-centerNumber of centersCountryStudy design1Wei L.20261,30670%30%NANoNoNo1ChinaRetrospective2Birtolo M. F.20261096544NANoNoNo1ItalyRetrospective3Hiyama A.20261239825NAYesNoNo1JapanRetrospective4Ho U-C.202621880%*20%*NAYesNoNo1Taiwan, ChinaRetrospective5Acosta-Batlle J.2025189377%*23%*313*NoNoNo1SpainRetrospective6Jiyoung Song2026383268115NAYesNoNo1South KoreaRetrospective7Li Y.20254,30518918061,608NoYesYes5ChinaRetrospective8Cheng Jia202516111348NAYesNoNo1ChinaRetrospective9Li Y.20251,167495169503NoYesYes4ChinaRetrospective10Park J.202575,53346,0486,49822,987NoYesYes>5South Korea, United States, Latin America, etc.Retrospective11Tang J.202517801,184296300YesNoNo1ChinaRetrospective12Ren Q.2025509279120110NoYesYes2ChinaRetrospective13Li Y.2025987561177249NoYesYes4ChinaProspective14Fanelli F.20253012416056NoYesYes2ItalyRetrospective15Du C.2025456248107101NoNoNo1ChinaRetrospective16Wu Y.20243,3122,3379754401*NoYesNo1ChinaRetrospective17Farid Gharehmohammadi2025907Unspecified, total 729178YesNoNo1United StatesRetrospective18Chengbin Huang20251,126581NA545NoYesYes4ChinaRetrospective19Qiye Cheng202430721493NANoNoNo1ChinaRetrospective20Fabio Galbusera2024429Unspecified, total 37386YesNoNo1SwitzerlandRetrospective21Heejun Park2024420UnspecifiedUnspecifiedUnspecifiedNoYesYes2South KoreaRetrospective22Jing Pan20241,6855301001,055YesYesNo1ChinaRetrospective23Xiaocong Lin2024930545143242NoYesYes3ChinaRetrospective24Sung Hye Kong20241709UnspecifiedUnspecified495YesYesYes2South KoreaRetrospective25Min Su Park20241,122831291NAYesNoNo1South KoreaRetrospective26Xiaoqing Yuan202322215567NAYesNoNo1ChinaRetrospective27Yu-Pin Chen202417301,430300NANoNoNo1Taiwan, ChinaRetrospective28Shigeng Wang20242,2741,660267347NoNoNo1ChinaRetrospective29Yu-Hsuan Lai2024412UnspecifiedUnspecifiedNANoNoNo1Taiwan, ChinaRetrospective30Shigeng Wang2023442239103100NoYesNo1ChinaRetrospective31Jing Pan20241,048530100418NoNoNo1ChinaRetrospective32Tao Peng20231,219703176340NoYesYes3ChinaRetrospective33Yaling Pan20231,175195698282NoYesYesUnknownChinaRetrospective34Keisuke Uemura2023857315542NAYesYesYes3JapanRetrospective35Xinyi Niu2023500300100100YesNoNo1ChinaRetrospective36Amara Tariq20234,4063,525881NAYesYesYes4United StatesRetrospective model development, prospective validation37Yung-Chieh Chen2023593Unspecified, total 197396YesYesNo1Taiwan, ChinaRetrospective38Jingnan Cui202327820870NAYesYesNo1ChinaRetrospective39Ronnie Sebro2022394274120NAYesNoNo1United StatesRetrospective40E Biamonte202224018060NAYesNoNo1ItalyRetrospective41Zhihao Xue2022133UnspecifiedUnspecifiedNAYesNoNo1ChinaRetrospective42Ronnie Sebro202219696100NAYesNoNo1United StatesRetrospective43Liyu Liu20222,2101768442NAYesNoYes2ChinaRetrospective44Miso Jang202214,1159,8253,0781,212NoYesNo1South KoreaRetrospective45Tomi Nissinen20213,523UnspecifiedUnspecified574YesYesYes2FinlandRetrospective model development, secondary analysis of prospective data for model validation46Yaling Pan202057416040374NoNoNo1ChinaRetrospective47Samir D Mehta201930724661NAYesNoNo1United StatesRetrospective48A Valentinitsch20191540.80.2NAYesNoNo1GermanyRetrospective49Chen-I Hsieh202141,89812,94023,3395,619YesYesNo1Taiwan, ChinaRetrospective50Qifei Dong20224,4163,413379669NoNoYes6United StatesRetrospective51İlkay Yıldız Potter202450545550NANoYesNo1United StatesRetrospective52Jun Tang202518801,184296400YesYesYes2ChinaRetrospective53Jaewon Kim202514,50210,1521,4502,900NoNoNo1South KoreaRetrospective54Amany M. Sarhan20241947*1558*389*NANoNoNoNA (public data)NA (multi-region)Retrospective55Xiaoling Zheng2024449344105NANoNoNo1ChinaRetrospective56Candan Güngör2025140UnspecifiedUnspecifiedUnspecifiedYesNoNo1TurkeyRetrospective57Fatih Erdem202414090%10%NAYesNoNo1TurkeyRetrospective*Represents image-level quantity; NA, not applicable. Characteristics of study design. *Represents image-level quantity; NA, not applicable. In terms of study design, retrospective analyses were overwhelmingly predominant. Only three studies incorporated prospective elements or validation (42, 65, 84), indicating that current evidence remains largely built on retrospective cohort designs. Population characteristics The included studies exhibited heterogeneity in population characteristics. Sample sizes varied widely, ranging from 109 to 75,524 participants, with a median of approximately 900. Notably, a study reported only the number of images rather than individual participants, reflecting inconsistencies in reporting practices (70). Most studies included both sexes, but females were consistently overrepresented, comprising approximately 60–65% of study populations on average. A small number of studies (n = 4) focused exclusively on women (31, 43, 58, 81), while one study included only male participants (79). This female predominance aligns with the higher prevalence of osteoporosis in women. Age inclusion criteria also varied considerably, with lower thresholds ranging from ≥18 to ≥65 years, although most studies adopted thresholds of ≥40 or ≥50 years. This variability reflects a lack of standardization in target populations and may influence model performance. In particular, inclusion of younger, lower-risk individuals may dilute disease prevalence and affect model discrimination, whereas higher thresholds may improve screening efficiency but could miss early-stage disease. Considerable variability was also observed in the clinical context of the study populations. Broadly, three categories could be identified: asymptomatic individuals undergoing routine health examinations, patients with specific underlying conditions (e.g., breast cancer (31), adrenal adenoma (34), degenerative spinal disorders (32, 49), or hip fractures (71)), and outpatient or emergency populations. This diversity reflects the opportunistic nature of screening, where imaging acquired for diverse clinical indications is repurposed for osteoporosis risk assessment. However, it also introduces important challenges, as differences in underlying disease status and clinical context may influence bone density distribution and imaging characteristics, potentially limiting model generalizability and necessitating domain adaptation across populations. Table 2 provides a detailed summary of these population characteristics. Table 2Summary of population characteristics.Study IDFirst authorFemaleMaleAge rangeAverage agePopulationGenderInclusion age threshold1Wei L.82348355–9463.96 ± 7.92Healthy individualsBoth sexes≥552Birtolo M. F.109027–8561.1 ± 11.3Breast cancer patientsFemales only>183Hiyama A.7449≥5071.9 ± 8.3Patients with degenerative spinal disordersBoth sexes≥504Ho U-C.17048≥5072.69 ± 9.96Lumbar spine disease patientsBoth sexes≥505Acosta-Batlle J.1706187Not mentioned72.3 ± 11.3Healthy individualsBoth sexesUnspecified (Average>72)6Jiyoung Song192191≥4059.8 ± 11.1Healthy individualsBoth sexes≥407Li Y.2,7921,51330–89Not mentionedHealthy individualsBoth sexes>308Cheng Jia679436–60Median age 53Patients with adrenal adenomaBoth sexesUnspecified (observed range 36–60)9Li Y.605562Not mentionedNot mentionedPatients undergoing LDCT for lung cancer screeningBoth sexesUnspecified10Park J.38,76736,75740–100Not mentionedHealth check population and clinical patientsBoth sexes≥4011Tang J.1,04673450–9969.46 ± 10.33Patients undergoing chest X-rayBoth sexes≥5012Ren Q.23727230–80Not mentionedPatients undergoing abdominal CTBoth sexes≥3013Li Y.59139630–89Not mentionedPatients undergoing LDCT for lung cancer screeningBoth sexes≥3014Fanelli F.3010≥50Not mentionedPostmenopausal womenFemales only≥5015Du C.23921750–9167.80 ± 8.18Patients undergoing low-dose abdominal CTBoth sexes≥5016Wu Y.1,5401772Not mentionedNot mentionedHealthy individualsBoth sexesUnspecified17Farid Gharehmohammadi767140≥5066.9 ± 9.0Healthy individualsBoth sexes≥5018Chengbin HuangUnspecifiedUnspecified≥50Not mentionedPatients undergoing chest CTBoth sexes≥5019Qiye Cheng86128Not mentionedNot mentionedPatients undergoing abdominal DEsCTBoth sexesUnspecified20Fabio Galbusera31311633–92Median age 72Patients with degenerative spinal disordersBoth sexesUnspecified (observed range 33–92)21Heejun Park260160Not mentioned60.3 ± 15.4Patients undergoing chest/abdominal/lumbar spine CTBoth sexesUnspecified22Jing Pan81087520–92Not mentionedHealthy individualsBoth sexes≥1823Xiaocong Lin647283≥18Not mentionedPatients undergoing chest CTBoth sexes≥1824Sung Hye Kong1,22948050–80Not mentionedPatients undergoing abdominal CTBoth sexes50–8025Min Su Park989133≥50Not mentionedPatients undergoing abdominopelvic CTBoth sexes≥5026Xiaoqing Yuan1487450–9266.4 ± 9.4Patients undergoing abdominal CTBoth sexes≥5027Yu-Pin Chen1,332398≥5072.4 ± 11.1Patients undergoing hip radiographyBoth sexes≥5028Shigeng Wang1,0091,265Not mentionedNot mentionedPatients undergoing chest CTBoth sexesUnspecified29Yu-Hsuan Lai4120Not mentioned61.5 ± 10.1Female breast cancer patientsFemales onlyUnspecified30Shigeng Wang178264Not mentioned61.86 ± 11.38Patients undergoing low-dose chest CT (LDCT)Both sexesUnspecified31Jing Pan44360520–9251.19 ± 14.35Healthy individualsBoth sexes≥1832Tao PengUnspecifiedUnspecifiedNot mentionedNot mentionedPatients undergoing opportunistic CT (chest, abdomen, lumbar spine) at multiple centersBoth sexesUnspecified33Yaling Pan40277332–88Not mentionedPatients undergoing chest LDCTBoth sexesUnspecified (observed range 32–88)34Keisuke Uemura668189Not mentioned70.0 ± 16.8Patients undergoing hip surgeryBoth sexesUnspecified35Xinyi NiuUnspecifiedUnspecified39–9066.03 ± 9.71General populationBoth sexesUnspecified (observed range 39–90)36Amara Tariq2,8781,528Not mentioned66.9 ± 9.2Patients undergoing abdominopelvic CTBoth sexesUnspecified37Yung-Chieh ChenUnspecifiedUnspecifiedNot mentionedNot mentionedHealthy individualsBoth sexesUnspecified38Jingnan CuiUnspecifiedUnspecifiedNot mentionedNot mentionedPostmenopausal women (spine); healthy and osteoporotic fracture patients (calcaneus)Spine: females only; calcaneus: unspecified≥50 (Spine)39Ronnie Sebro28610850–8762.1 ± 7.4Patients undergoing abdominopelvic CTBoth sexes≥5040E Biamonte110130≥1860.4 ± 15.4Patients undergoing lumbar spine CTBoth sexes≥1841Zhihao Xue924131–9465.45 ± 9.82Patients with low back pain undergoing lumbar spine CTBoth sexesUnspecified (observed range 31–94)42Ronnie Sebro1732350–8864.9 ± 8.7Patients undergoing wrist CTBoth sexes≥5043Liyu LiuUnspecifiedUnspecified>40Not mentionedHealth check population and clinical patientsBoth sexes>4044Miso Jang11,8112,30440–90Not mentionedHealth check population (training/internal test) and clinical patients (external test)Both sexes>4045Tomi Nissinen3,40811520–96Not mentionedPostmenopausal women and clinical patientsBoth sexesUnspecified46Yaling PanUnspecifiedUnspecified50–8862.6 ± 7.6Patients undergoing LDCTBoth sexes≥5047Samir D Mehta202105Not mentionedNot mentionedPatients undergoing DXABoth sexesUnspecified (observed range 67–71)48A Valentinitsch5110339–8864.6 ± 8.5Cancer patientsBoth sexes>3849Chen-I Hsieh18,4664,87340–90Not mentionedPatients undergoing pelvic/spine X-rayBoth sexes40–9050Qifei Dong04,416≥6573.7Community-dwelling older menMales only≥6551İlkay Yıldız Potter18613018–100Not mentionedEmergency or outpatient patientsBoth sexesUnspecified (observed range 18–100)52Jun Tang1,104776≥50Not mentionedClinical patientsBoth sexes≥5053Jaewon Kim14,502020–8855.97 ± 9.69Female health check populationFemales onlyUnspecified (observed range 20–88)54Amany M. SarhanUnspecifiedUnspecifiedNot mentionedNot mentionedPatients undergoing knee radiographyUnspecifiedUnspecified55Xiaoling Zheng300149Not mentionedNot mentionedHip fracture patientsBoth sexesUnspecified56Candan Güngör1241633–91Not mentionedClinical patientsBoth sexesUnspecified (observed range 33–91)57Fatih Erdem1328Not mentioned65.32 ± 8.50Clinical patientsBoth sexesUnspecifiedDXA, dual-energy X-ray absorptiometry. Summary of population characteristics. DXA, dual-energy X-ray absorptiometry. Overall, the marked heterogeneity in study design, sample size, demographic composition, and clinical context should be carefully considered when interpreting subsequent findings on imaging modalities and AI model performance. Modeling and imaging characteristics: predominance of automated deep learning Adoption of automated imaging ROI Table 3 summarizes the key ROI characteristics of the included studies. CT served as the primary data source for AI-driven opportunistic osteoporosis screening, being employed in 38 studies and encompassing a range of acquisition protocols, including conventional, low-dose, non-contrast, and contrast-enhanced multi-detector CT. X-ray-based approaches were also frequently utilized (n = 15). In contrast, MRI was only rarely applied (n = 3) (33, 49, 86), consistent with its limited role in the quantitative assessment of bone mineral density. Despite its status as the clinical reference standard, DXA was incorporated into opportunistic screening workflows in 2 studies (78, 84). In these cases, DXA data were analyzed in an automated manner without requiring additional imaging, reflecting efforts to extend its utility beyond traditional diagnostic settings. Table 3Summary of region of interest characteristics.Study IDFirst authorModalityROIROI categoryROI segmentation1Wei L.CTT10, T11, T12, and L1 vertebraeSpinal vertebraeAI-based automatic segmentation2Birtolo M. F.CTL1–L5 vertebraeSpinal vertebraeAI-based automatic segmentation3Hiyama A.CTL1–L4 vertebraeSpinal vertebraeManual delineation4Ho U-C.MRIL1–L4 vertebraeSpinal vertebraeManual delineation5Acosta-Batlle J.X-rayProximal femurHip/proximal femurAutomatic segmentation and manual delineation6Jiyoung SongCTThoracolumbar vertebraeSpinal vertebraeAI-based automatic segmentation7Li Y.LDCTT12–L2 vertebraeSpinal vertebraeAI-based automatic segmentation8Cheng JiaCTThoracolumbar vertebraeSpinal vertebraeManual delineation9Li Y.LDCTT12–L2 vertebraeSpinal vertebraeAI-based automatic segmentation10Park J.X-rayChestWhole image/non-precise segmentationAI-based automatic segmentation11Tang J.X-rayChestWhole image/non-precise segmentationAI-based automatic segmentation12Ren Q.CTL1–L3 vertebraeSpinal vertebraeManual delineation13Li Y.LDCTT12–L2 vertebraeSpinal vertebraeAI-based automatic segmentation14Fanelli F.X-rayPosterior mandible regionWhole image/non-precise segmentationManual delineation15Du C.LDCTProximal femurHip/proximal femurAI-based automatic segmentation16Wu Y.LDCTT12–L2 vertebraeSpinal vertebraeAI-based automatic segmentation17Farid GharehmohammadiX-rayFeet or ankleWhole image/non-precise segmentationAI-based automatic segmentation18Chengbin HuangCTT12 vertebra and surrounding skeletal musclesSpinal vertebraeManual delineation19Qiye ChengCTL1–L3 vertebraeSpinal vertebraeManual delineation20Fabio GalbuseraMRI, X-rayL1–L4 vertebraeSpinal vertebraeAI-based automatic segmentation21Heejun ParkCTL1 and L2 vertebraeSpinal vertebraeAI-based automatic segmentation22Jing PanCTL1 and L2 vertebraeSpinal vertebraeAI-based automatic segmentation23Xiaocong LinCTT11 or T12 vertebraSpinal vertebraeSemi-Automatic segmentation24Sung Hye KongCTT12–L4 vertebraeSpinal vertebraeAI-based automatic segmentation25Min Su ParkCTProximal femurHip/proximal femurAI-based automatic segmentation26Xiaoqing YuanCTLeft proximal femurHip/proximal femurAI-based automatic segmentation27Yu-Pin ChenX-rayProximal femurHip/proximal femurAI-based automatic segmentation28Shigeng WangCTT10–T12 vertebraeSpinal vertebraeAI-based automatic segmentation29Yu-Hsuan LaiCTL1 vertebraSpinal vertebraeManual delineation30Shigeng WangLDCTT10–T12 vertebraeSpinal vertebraeAI-based automatic segmentation31Jing PanCTL1 and L2 vertebraeSpinal vertebraeAI-based automatic segmentation32Tao PengCTT12–L2 or L1–L3 vertebraeSpinal vertebraeAI-based automatic segmentation33Yaling PanLDCTT11–L2 vertebraeSpinal vertebraeAI-based automatic segmentation34Keisuke UemuraCTProximal femurHip/proximal femurAI-based automatic segmentation35Xinyi NiuLDCTT12–L2 vertebraeSpinal vertebraeAI-based automatic segmentation36Amara TariqCTL3 vertebra, hip, and lumbar regionMulti-site/combinationAI-based automatic segmentation37Yung-Chieh ChenLDCTL1–L4 vertebraeSpinal vertebraeAI-based automatic segmentation38Jingnan CuiX-rayL4 vertebra and calcaneusMulti-site/combinationSemi-Automatic segmentation39Ronnie SebroCTL1–L4 vertebrae, pelvis and proximal femurMulti-site/combinationSemi-Automatic segmentation40E BiamonteCTL1–L5 vertebraeSpinal vertebraeManual delineation41Zhihao XueCTL1–L4 vertebraeSpinal vertebraeAI-based automatic segmentation42Ronnie SebroCTRadius, ulna, carpals, metacarpalsWhole image/non-precise segmentationManual delineation43Liyu LiuCTL1–L4 vertebraeSpinal vertebraeManual delineation44Miso JangX-rayChestWhole image/non-precise segmentationAI-based automatic segmentation45Tomi NissinenDXAL1–L4 vertebraeSpinal vertebraeAI-based automatic segmentation46Yaling PanLDCTT12–L2 vertebraeSpinal vertebraeAI-based automatic segmentation47Samir D MehtaDXAL1–L4 vertebraeSpinal vertebraeNo segmentation (using DEXA report data)48A ValentinitschMDCTT12–L4 vertebraeSpinal vertebraeAI-based automatic segmentation49Chen-I HsiehX-rayL1–L4 vertebrae and proximal femurMulti-site/combinationAI-based automatic segmentation50Qifei DongX-raySingle spine vertebraSpinal vertebraeManual delineation51İlkay Yıldız PotterX-rayhipWhole image/non-precise segmentationNo segmentation, whole image input52Jun TangX-rayChestWhole image/non-precise segmentationNo segmentation, whole image input53Jaewon KimX-rayChestWhole image/non-precise segmentationNo segmentation, whole image input54Amany M. SarhanX-raykneeWhole image/non-precise segmentationNo segmentation, whole image input55Xiaoling ZhengCTT11 or T12 vertebraSpinal vertebraeSemi-Automatic segmentation56Candan GüngörCTClivusClivusSemi-Automatic segmentation57Fatih ErdemMRIlumbarSpinal vertebraeManual delineationCT, Computed Tomography; LDCT, Low-Dose Computed Tomography; MRI, Magnetic Resonance Imaging; ROI, Region of Interest; AI, Artificial intelligence; DXA, dual-energy X-ray absorptiometry. Summary of region of interest characteristics. CT, Computed Tomography; LDCT, Low-Dose Computed Tomography; MRI, Magnetic Resonance Imaging; ROI, Region of Interest; AI, Artificial intelligence; DXA, dual-energy X-ray absorptiometry. From an anatomical perspective, the spine emerged as the predominant ROI, featuring in 35 studies. Analyses typically focused on vertebral levels such as L1-L4 or T12-L2, given their strong correlation with bone mineral density and their suitability for standardized assessment. Other studies targeted the hip or proximal femur (n = 6) (30, 44, 54–56, 63) or adopted multisite strategies integrating multiple skeletal regions (n = 4). More recently, opportunistic screening has been extended to the craniofacial region, with two included studies leveraging clivus-based CT radiomic features (72) and mandibular radiographs (43), respectively. These emerging anatomical targets carry considerable clinical relevance, as head and maxillofacial imaging is routinely performed in emergency, dental, and geriatric care, generating a vast, currently underutilized resource for bone health assessment. While the clivus approach exploits feature already present in standard head CTs, mandibular radiographs offer a complementary, low-radiation modality that could further broaden population-level screening reach. Notably, ten studies dispensed with predefined ROIs altogether, instead leveraging whole-image inputs. Regarding ROI extraction, deep learning-based automatic segmentation was the most common approach (n = 34), enabling scalable and fully automated pipelines. Manual (n = 12) (32–34, 41, 43, 47, 48, 58, 74, 77, 79, 82) and semi-automated (n = 4) (52, 67, 71, 76) methods were less frequent. Studies that used whole image inputs did not require explicit ROI segmentation, further simplifying preprocessing workflows. Recent investigations have increasingly emphasized fully automated segmentation strategies, transitioning from manual delineation to deep learning-based methods (e.g., nnU-Net, DeepLab, DeepCatch), thereby substantially reducing labor demands. While such approaches may better align with real-world clinical deployment, they may also diminish anatomical interpretability. Together, these patterns underscore a broader shift toward automation and streamlined model inputs, a trajectory that extends to subsequent stages of feature extraction and model development. Deep learning as the dominant modeling framework Table 4 summarizes the feature extraction methods and prediction algorithms. Feature extraction strategies could be broadly categorized into three groups. Deep learning-based approaches were the most prevalent (n = 29), enabling end-to-end representation learning directly from imaging data. Handcrafted radiomics features were employed in 18 studies, typically involving the extraction of texture and shape descriptors followed by feature selection procedures. A smaller subset of studies relied on conventional CT attenuation metrics, such as Hounsfield units (HU) (n = 5). In addition, two studies combined deep learning-derived features with radiomics, while one study utilized quantitative outputs from DXA as numerical inputs. Table 4Characteristics of prediction models.Study IDFirst authorFeature extractionPrediction model algorithmModel type (Combined imaging and clinical vs. imaging-only)Model comparison performed1Wei L.Deep Learning FeaturesConventional statistical modelImaging-onlyNo2Birtolo M. F.Handcrafted RadiomicsConventional statistical model and machine learningCombined modelYes3Hiyama A.HU MeasurementsConventional statistical model and machine learningCombined modelNo4Ho U-C.Deep Learning FeaturesDeep learning modelImaging-onlyNo5Acosta-Batlle J.Deep Learning FeaturesDeep learning modelImaging-onlyNo6Jiyoung SongDeep Learning FeaturesTraditional machine learningCombined modelYes7Li Y.Deep Learning FeaturesDeep learning modelImaging-onlyNo8Cheng JiaHandcrafted RadiomicsConventional statistical modelCombined modelYes9Li Y.Deep Learning FeaturesDeep learning modelImaging-onlyNo10Park J.Deep Learning FeaturesDeep learning modelImaging-onlyNo11Tang J.Deep Learning FeaturesDeep learning modelCombined modelYes12Ren Q.Handcrafted RadiomicsTraditional machine learningImaging-onlyNo13Li Y.Deep Learning FeaturesDeep learning modelImaging-onlyNo14Fanelli F.Deep Learning Features and Handcrafted RadiomicsDeep learning modelImaging-onlyNo15Du C.Handcrafted RadiomicsTraditional machine learningImaging-onlyNo16Wu Y.Deep Learning FeaturesDeep learning modelImaging-onlyNo17Farid GharehmohammadiDeep Learning FeaturesDeep learning modelImaging-onlyNo18Chengbin HuangDeep Learning FeaturesDeep learning modelImaging-onlyYes19Qiye ChengHandcrafted RadiomicsTraditional machine learningImaging-onlyNo20Fabio GalbuseraHandcrafted RadiomicsDeep learning and traditional machine learningCombined modelYes21Heejun ParkDeep Learning FeaturesDeep learning modelImaging-onlyNo22Jing PanDeep Learning Features and Handcrafted RadiomicsDeep learning modelCombined modelYes23Xiaocong LinHandcrafted RadiomicsTraditional machine learningCombined modelYes24Sung Hye KongDeep Learning FeaturesDeep learning modelCombined modelYes25Min Su ParkHandcrafted RadiomicsTraditional machine learningCombined modelNo26Xiaoqing YuanHandcrafted RadiomicsConventional statistical modelImaging-onlyNo27Yu-Pin ChenDeep Learning FeaturesDeep learning modelCombined modelNo28Shigeng WangDeep Learning FeaturesDeep learning modelImaging-onlyNo29Yu-Hsuan LaiHandcrafted RadiomicsConventional statistical modelCombined modelYes30Shigeng WangHandcrafted RadiomicsTraditional machine learningImaging-onlyNo31Jing PanDeep Learning FeaturesDeep learning modelImaging-onlyNo32Tao PengDeep Learning FeaturesDeep learning modelImaging-onlyNo33Yaling PanHU MeasurementsConventional statistical modelImaging-onlyNo34Keisuke UemuraHU MeasurementsDeep learning modelImaging-onlyNo35Xinyi NiuDeep Learning FeaturesDeep learning modelImaging-onlyNo36Amara TariqDeep Learning FeaturesDeep learning, traditional machine learning and conventional statistical modelCombined modelYes37Yung-Chieh ChenHandcrafted RadiomicsTraditional machine learningImaging-onlyNo38Jingnan CuiHandcrafted RadiomicsTraditional machine learningImaging-onlyNo39Ronnie SebroHU MeasurementsConventional statistical model and machine learningCombined modelNo40E BiamonteHandcrafted RadiomicsTraditional machine learningImaging-onlyNo41Zhihao XueHandcrafted RadiomicsTraditional machine learningImaging-onlyNo42Ronnie SebroHU MeasurementsTraditional machine learningImaging-onlyYes43Liyu LiuHandcrafted RadiomicsConventional statistical model and machine learningCombined modelYes44Miso JangDeep Learning FeaturesDeep learning modelImaging-onlyNo45Tomi NissinenDeep Learning FeaturesDeep learning modelImaging-onlyNo46Yaling PanDeep Learning FeaturesDeep learning modelImaging-onlyNo47Samir D MehtaDXA-derived Numerical FeaturesTraditional machine learningImaging-onlyNo48A ValentinitschHandcrafted RadiomicsTraditional machine learningImaging-onlyNo49Chen-I HsiehDeep Learning FeaturesDeep learning modelImaging-onlyNo50Qifei DongDeep Learning FeaturesDeep learning modelImaging-onlyNo51İlkay Yıldız PotterDeep Learning FeaturesDeep learning modelImaging-onlyNo52Jun TangDeep Learning FeaturesDeep learning modelImaging-onlyNo53Jaewon KimDeep Learning FeaturesDeep learning modelImaging-onlyNo54Amany M. SarhanDeep Learning FeaturesDeep learning modelImaging-onlyNo55Xiaoling ZhengHandcrafted RadiomicsTraditional machine learningCombined modelYes56Candan GüngörHandcrafted RadiomicsTraditional machine learningImaging-onlyNo57Fatih ErdemHandcrafted RadiomicsTraditional machine learningImaging-onlyNoModel type: imaging-only, model built with imaging features alone; combined, model built with imaging plus clinical features. Model comparison performed: yes/no indicates whether the study directly compared the predictive performance of the imaging-only and combined models. DXA, dual-energy X-ray absorptiometry. Characteristics of prediction models. Model type: imaging-only, model built with imaging features alone; combined, model built with imaging plus clinical features. Model comparison performed: yes/no indicates whether the study directly compared the predictive performance of the imaging-only and combined models. DXA, dual-energy X-ray absorptiometry. In terms of predictive frameworks, deep learning models were most frequently adopted (n = 29), followed by traditional machine learning approaches (n = 15), including support vector machines (SVM), random forests (RF), and logistic regression (LR). The remaining studies employed conventional statistical models or hybrid methodologies. A clear temporal trend was observed: earlier studies (2019–2022) more commonly relied on HU-based features or radiomics coupled with machine learning, whereas studies published from 2023 onward predominantly adopted deep learning architectures. Widely used deep learning architectures included convolutional neural networks (e.g., ResNet, DenseNet, EfficientNet, Inception, and VGG) and segmentation frameworks such as U-Net-based models. These architectures often supported end-to-end pipelines, spanning from ROI segmentation to downstream tasks such as bone mineral density estimation, osteoporosis classification, and fracture risk prediction. Notably, several studies published in 2025 reported fine-tuning of large pretrained models, including OpenCLIP, DINOv2, and CheXagent, highlighting an emerging shift toward foundation model-based paradigms (69, 70, 81). Compared with radiomics-based approaches, deep learning models generally offer improved reproducibility and reduced reliance on manual intervention. Nonetheless, radiomics retains value in terms of interpretability, as extracted features can be directly mapped to specific image characteristics. Hybrid strategies, such as combining deep learning-based segmentation with radiomics-driven classification, were also observed, aiming to balance automation with interpretability. Overall, the field is transitioning toward increasingly automated modeling frameworks, with deep learning assuming a central role in both feature extraction and predictive modeling. The underestimated potential of clinical features in osteoporosis screening The majority of included studies developed imaging-only models (n = 39), whereas a smaller subset (n = 16) incorporated both imaging and clinical variables. Notably, no study relied exclusively on clinical features. Among studies adopting a multimodal approach, commonly included clinical variables comprised age, sex, body mass index, and laboratory parameters. Of these, 13 studies explicitly compared the performance of imaging-only models with that of combined models, yielding largely consistent findings (Table 4). Overall, the integration of clinical data conferred modest yet measurable improvements in predictive performance, as reflected by increases in metrics such as the area under the curve (AUC) and overall accuracy. However, several studies reported negligible or no incremental benefit from including clinical variables (38, 52, 71), particularly in scenarios where imaging models had already captured rich, high-dimensional representations of the underlying pathology. For example, models leveraging vertebral imaging and body composition features demonstrated strong predictive performance without appreciable gains from additional clinical inputs (38). The model comparison information is reported in Table 4. These findings should not be taken to imply that clinical features lack predictive value. A very recent study demonstrated that a model relying solely on 35 routine clinical variables achieved an accuracy of 83.94% for opportunistic osteoporosis screening across three hospitals (87). The model combined Transformer, Mamba, and a FeatureBake Block with a CACM module, and SHAP analyses identified hemoglobin, age, and weight as the most influential predictors, linking model decisions to clear physiological rationales. This evidence highlights that, even without imaging, clinical variables can carry substantial discriminative power. The limited incremental gains observed in earlier imaging-dominant settings may therefore reflect ceiling effects rather than an intrinsic weakness of clinical data. Collectively, these findings suggest that the modest gains from adding clinical variables in prior imaging-based studies likely reflect ceiling effects of high-performance imaging models, rather than an inherent lack of predictive value in clinical data. Quality assessment: deficits in external validation as a primary source of bias To systematically evaluate the quality of the included literature, we employed two frameworks: the phase classification criteria and the Prediction Model Risk of Bias Assessment Tool (PROBAST). The phase classification criteria stratify studies according to sample size, study design (retrospective vs. prospective), and validation status (none, internal, or external). Given that all studies enrolled more than 100 participants and were predominantly retrospective in design, 55 were classified as Phase II, whereas the remaining two studies involving prospective data collection or validation were accordingly designated Phase III (Table 5). Inter-reviewer agreement was 98.25% (56/57), with a Kappa of 0.66. The modest Kappa value despite near-perfect observed agreement is attributable to the severe class imbalance in the data, which inflates the expected agreement and exerts a strong downward correction on Kappa. The wide confidence interval further reflects the small number of studies and the extreme prevalence of Phase II ratings. Nevertheless, the highly observed agreement indicates that inter-reviewer reliability is practically acceptable. Table 5Summary of the quality of included studies.Study IDFirst authorPhaseROB1Wei L.IIHigh2Birtolo M. F.IIHigh3Hiyama A.IIHigh4Ho U-C.IIHigh5Acosta-Batlle J.IIHigh6Jiyoung SongIIHigh7Li Y.IILow8Cheng JiaIIHigh9Li Y.IILow10Park J.IILow11Tang J.IIHigh12Ren Q.IILow13Li Y.IILow14Fanelli F.IILow15Du C.IIHigh16Wu Y.IILow17Farid GharehmohammadiIIHigh18Chengbin HuangIILow19Qiye ChengIIHigh20Fabio GalbuseraIIHigh21Heejun ParkIILow22Jing PanIILow23Xiaocong LinIILow24Sung Hye KongIILow25Min Su ParkIIHigh26Xiaoqing YuanIIHigh27Yu-Pin ChenIIHigh28Shigeng WangIIHigh29Yu-Hsuan LaiIIHigh30Shigeng WangIILow31Jing PanIIHigh32Tao PengIILow33Yaling PanIILow34Keisuke UemuraIILow35Xinyi NiuIIHigh36Amara TariqIIILow37Yung-Chieh ChenIILow38Jingnan CuiIILow39Ronnie SebroIIHigh40E BiamonteIIHigh41Zhihao XueIIHigh42Ronnie SebroIIHigh43Liyu LiuIIHigh44Miso JangIILow45Tomi NissinenIIILow46Yaling PanIIHigh47Samir D MehtaIIHigh48A ValentinitschIIHigh49Chen-I HsiehIILow50Qifei DongIIHigh51İlkay Yıldız PotterIILow52Jun TangIILow53Jaewon KimIIHigh54Amany M. SarhanIIHigh55Xiaoling ZhengIIHigh56Candan GüngörIIHigh57Fatih ErdemIIHighPhase, phase classification criteria; ROB, risk of bias. Summary of the quality of included studies. Phase, phase classification criteria; ROB, risk of bias. Risk of bias was further assessed using PROBAST across four domains: Participants, Predictors, Outcome, and Analysis. As this review focuses on model development studies, the absence of external validation is, by PROBAST standards, indicative of high risk of bias. Accordingly, 25 studies were judged to be at high risk. Domain-specific assessments are summarized in Table 5 and Figure 2. Evaluation of risk of bias of included studies. Validation designs varied substantially across studies, particularly in data partitioning, use of resampling methods, and implementation of external validation. Most studies adopted simple train-test splits, typically using ratios such as 70/30 or 80/20. In several cases, validation and testing were not clearly distinguished, and some lacked an independent test set, relying instead on internal validation or cross-validation. Approximately half of the studies employed cross-validation, primarily in the context of limited sample sizes, as an attempt to enhance robustness through resampling. While this approach can reduce overfitting, it does not replace the need for independent external validation. Only 20 studies performed external validation. Among these, a small proportion used truly independent datasets from different institutions or populations, including a limited number of multicenter studies. In many cases, external validation was conducted within similar geographic or institutional settings, which may limit the assessment of model performance across diverse populations and imaging conditions. Truly international and multi-ethnic validation was rare, with only isolated examples (37). Sample size was not consistently associated with a stronger study design. Although some large studies (>10,000 participants), such as Park (2025) (37) and Jang (2022) (80), included external validation, most studies with moderate sample sizes (1,000-5,000) did not, and smaller studies (<500) generally relied on internal validation alone. It is also important to note that large retrospective, single-center studies may still be affected by selection bias and site-specific imaging characteristics. Overall, current studies show increasing but still limited use of external and multicenter validation. Notably, this review primarily focuses on model development studies. Compared with model validation studies, model development studies are inherently more difficult to validate across multiple centers due to external resource constraints. This is an intrinsic limitation of this research type. These factors should be considered when interpreting reported model performance, which is examined in the following section. Model performance: variability across tasks and clinical endpoints To enhance transparency and clinical applicability, we performed descriptive stratification of the 57 included studies by predicted outcome (osteoporosis diagnosis, osteopenia, and fracture risk), imaging modality, and model type (Table 6). Overall, CT was the dominant imaging modality (particularly for osteoporosis and osteopenia prediction), while deep learning approaches predominated across all outcome categories. Fracture risk prediction had the fewest studies and generally lower performance. Detailed stratification is presented in Table 6. Table 6Descriptive stratification of included studies and model performance by imaging modality and model type.Predicted outcomeImaging modalityModel typeNo. of studiesNo. with external validation (%)Median AUC (Range)OsteoporosisCTDeep Learning137 (54%)0.94 (0.82–1.0)Traditional ML/Hybrid156 (40%)0.91 (0.79–0.99)MRIDeep Learning10 (0%)0.83Traditional ML/Hybrid20 (0%)0.86 (0.8–0.91)X-rayDeep Learning94 (44%)0.93 (0.81–0.99)Traditional ML/Hybrid22 (100%)0.78 (0.77–0.79)OsteopeniaCTDeep Learning138 (62%)0.94 (0.69–1)Traditional ML/Hybrid114 (36%)0.94 (0.81–0.98)MRITraditional ML/Hybrid10 (0%)0.82X-rayDeep Learning21 (50%)0.91 (0.87–0.95)Traditional ML/Hybrid11 (100%)0.72Fracture RiskCTDeep Learning11 (100%)0.83Traditional ML/Hybrid30 (0%)0.83 (0.79–0.88)DXADXA-derived Numerical Features21 (50%)0.77 (0.63–0.91)X-rayDeep Learning32 (67%)0.94 (0.82–0.99)AUC, area under the curve; CT, computed tomography; MRI, magnetic resonance imaging; DXA, dual-energy X-ray absorptiometry; ML, machine learning. Descriptive stratification of included studies and model performance by imaging modality and model type. AUC, area under the curve; CT, computed tomography; MRI, magnetic resonance imaging; DXA, dual-energy X-ray absorptiometry; ML, machine learning. Performance variability in osteoporosis prediction A total of 42 studies reported performance for osteoporosis prediction, comprising 45 models, as one study constructed separate models based on the T10, T11, T12, and L1 vertebrae (39). To compare the performance of these models on the same scale, we extracted the AUC values. Overall, model discrimination varied substantially, with AUC values ranging from 0.767 to 1.000. Notably, eight models achieved an AUC ≥ 0.98, including three reporting perfect discrimination (AUC = 1.000) (35, 36, 42). While these three studies employed patient-level splitting with external validation sets drawn from independent centers and scanners, perfect AUC values on external data remain highly implausible and should not be attributed merely to cohort size. A more likely explanation is some form of data leakage that cannot be fully discerned from the published methods. This concern is reinforced by the observation in Ref (33) that AUC values of 1.000 were achieved across all test sets, both internal and external, a pattern difficult to reconcile with genuine generalization. Accordingly, these perfect scores warrant extreme caution in interpretation. Performance varied according to imaging modality and modeling strategy. As shown in Table 6, CT-based models constituted the majority of osteoporosis prediction studies, with deep learning models achieving a median AUC of 0.94 (range 0.82–1.00) and traditional machine learning (ML)/hybrid models a median of 0.91 (0.79–0.99). MRI-based models remained limited (n = 2), yielding more modest AUCs (0.80–0.86). Radiography (X-ray) studies showed strong performance in deep learning models (median AUC 0.93), though with fewer studies overall. The two MRI-based models yielded relatively modest AUCs (0.80 and 0.83) (33, 49). However, recent work suggests that functional MRI sequences may offer a more promising direction. Erdem et al. (2024) applied radiomic feature extraction to ADC maps of lumbar vertebrae and combined the selected features with machine learning classifiers (86). A Naive Bayes model trained on five wavelet- and Laplacian-of-Gaussian-derived features achieved an AUC of 0.913, with 90.7% accuracy and 95% specificity. Because ADC values reflect microarchitectural changes in bone marrow, such as shifts in fat and water content, functional sequences like DWI and ADC mapping may capture information that conventional T1- and T2-weighted images cannot. Although this evidence comes from a single study with a modest sample size and requires external validation, it indicates that advanced MRI sequences represent a viable pathway for AI-driven opportunistic osteoporosis screening and merit further investigation (88). Models using whole-image inputs without predefined regions of interest achieved strong performance (AUC: 0.82–0.99), albeit with reduced interpretability. In terms of feature extraction approaches, the deep learning-based approach (n = 27) demonstrated better performance (median AUC = 0.93) compared with handcrafted radiomics features (n = 12, median AUC = 0.90) and models based on Hounsfield unit measurements (n = 4, median AUC = 0.89). Task-dependent performance in osteopenia prediction A total of 28 studies developed models for osteopenia prediction, comprising 31 models, including four vertebra-specific models (T10, T11, T12, and L1) from a single study (39). Overall discrimination was highly variable (AUC: 0.69–1.000), with a lower performance bound than that observed for osteoporosis prediction (0.69 vs. 0.767), indicating reduced separability for this intermediate phenotype. Despite this variability, a subset of models demonstrated excellent performance, with seven achieving AUC ≥ 0.98, including four with perfect discrimination (AUC = 1.000) (35, 36, 42). Stratification by imaging modality (Table 6) revealed that CT remained the most common modality for osteopenia prediction as well. Deep learning models on CT achieved a median AUC of 0.94 (0.69–1.00), while traditional ML/hybrid models showed comparable performance (median 0.94). Radiography-based deep learning models also performed well (AUC 0.91), although data for MRI and other modalities were sparse. Heterogeneity in task design may partly explain this variability. Some studies adopted binary classification (normal vs. osteopenia), whereas others implemented multiclass classification (normal/osteopenia/osteoporosis), the latter representing a more clinically relevant yet inherently more challenging task. Notably, the lowest AUC (0.69) was reported in a multiclass model based on the L1 vertebra, using DXA-defined thresholds (−2.5 < T-score <= − 1.0), underscoring the difficulty of discriminating osteopenia from adjacent categories. Collectively, these findings suggest that osteopenia prediction is sensitive to task design and remains inherently more challenging despite the presence of high-performing models. Persistent challenges in fracture risk prediction Nine studies addressed fracture risk prediction (31, 53, 68, 78, 79, 82, 84, 85, 89), encompassing heterogeneous endpoints, including binary classification of osteoporotic vertebral compression fractures, discrimination of moderate-to-severe fractures from normal or mild cases, and site-specific outcomes (e.g., lumbar spine, proximal femur). Across these tasks, model performance was both more variable and consistently lower than that observed for bone density-based tasks, with AUCs ranging from 0.63 to 0.99. The lower performance bound, compared with osteoporosis (0.767–1.000) and osteopenia prediction (0.69–1.000), underscores the intrinsic complexity of fracture risk, which depends on multiple factors beyond bone mineral density, including fall propensity and skeletal structure. The lowest AUC (0.63) was reported in a Finnish study using DXA with prospective validation, representing the only model evaluated in a prospective cohort (84). In contrast, the highest AUC (0.99) was observed in a U. S.-based, multicenter X-ray study restricted to male participants but without external validation. Among studies with external validation (n = 4) (53, 68, 84, 89), the AUC ranged from 0.63 to 0.94, with a median of approximately 0.82. In contrast, studies without external validation (n = 5) reported a higher median AUC of approximately 0.88 (range: 0.79–0.99). This pattern suggests that the absence of external validation may lead to overestimation of model performance. Descriptive stratification highlighted the relative scarcity of fracture risk studies (Table 6). CT-based deep learning models (n = 1) achieved an AUC of 0.83, while traditional ML/hybrid models on CT showed a median AUC of 0.83 (0.79–0.88). X-ray-based deep learning models (n = 3) reported a higher median AUC of 0.94, but with limited external validation. Across tasks, fracture prediction demonstrated lower median discrimination (AUC = 0.83) compared with osteoporosis (0.93) and osteopenia prediction (0.93), with a broader lower tail. Descriptive stratification by imaging modality and model type (Table 6) further highlights these differences. Deep learning approaches generally achieved high median AUCs in osteoporosis (0.94 for CT) and osteopenia prediction, while fracture risk studies were relatively scarce. These findings indicate that while imaging-based models perform well for bone density classification, fracture risk prediction remains substantially more challenging. CT-based opportunistic screening currently dominates the literature, with limited evidence available for MRI and DXA modalities. Clinicians should interpret reported performance metrics in the context of the specific imaging modality, predicted outcome, and model type. Future high-quality studies with robust external validation across diverse modalities and fracture-focused endpoints are urgently needed to bridge these gaps. Sources of heterogeneity: drivers of study-level variation and performance variability To identify sources of heterogeneity, we performed principal component analysis on 12 variables extracted from 57 studies, including reference standard, AUC, participant number, external validation, multicenter design, number of centers, country, study design, gender, ROI category, modality, and image-based extraction strategy. Because osteoporosis and osteopenia represent distinct prediction tasks, analyses were conducted separately (Figure 3). Principal component analysis (PCA) of study-level heterogeneity. (A) PCA based on studies predicting osteoporosis. (B) PCA based on studies predicting osteopenia. The PCA was conducted using 12 key study-level variables from the included studies. The first two principal components (PC1 and PC2) are displayed. Red arrows represent the loading vectors of individual variables, indicating the direction and magnitude of their contribution to variance in the PCA space. The relative orientation and length of the arrows reflect correlations among variables and their influence on study heterogeneity. For simplicity, only the variables with the largest contributions to PC1 and PC2 are shown. In osteoporosis prediction, the first two principal components explained 49% of the variance (PC1: 25.9%; PC2: 23.1%). PC1 was mainly associated with the number of centers (loading = 0.67), AUC (loading = 0.40), and participant number (loading = −0.32), while PC2 was driven by participant number (loading = −0.70) and AUC (loading = 0.40). In osteopenia prediction, the first two components explained 52.2% of the variance (PC1: 29.3%; PC2: 22.9%). PC1 was primarily associated with AUC (loading = 0.71), participant number (loading = −0.33), and multicenter design (loading = −0.28), and PC2 with participant number (loading = −0.79), AUC (loading = 0.29), and reference standard (loading = 0.24). Across both tasks, smaller studies tended to report higher AUC. This pattern likely reflects multiple converging mechanisms rather than a single cause. Smaller studies are more susceptible to overfitting, as models trained on limited data may capture noise rather than generalizable signal. They may also be affected by selection bias, where the study cohort is more homogeneous than the target population, artificially inflating performance. Single-center studies generally showed higher performance than multicenter studies, a finding that may be explained by institution-specific acquisition protocols and population characteristics that reduce data variability and make the prediction task appear easier than it would be in a heterogeneous real-world setting. In multicenter studies, the inclusion of diverse scanners, protocols, and patient demographics introduces greater variability that challenges model generalization, resulting in slightly lower but more realistic AUC estimates. Overall, these results indicate that participant number, multicenter design, and the number of centers is key structural factors that contribute to the observed heterogeneity in reported model performance, and that the higher AUCs in smaller, single-center studies should be interpreted with appropriate caution. Discussion This systematic review synthesizes evidence from 57 studies published between 2019 and 2026, documenting the rapid development of AI-based opportunistic screening for osteoporosis. To our knowledge, this is the first comprehensive synthesis to characterize methodological heterogeneity, modeling trends, and performance variability in this field. The findings provide several insights relevant to clinical translation and future study design. The marked increase in publications since 2023, with nearly 80% of included studies published in the last 3 years, reflects a rapid expansion of the field. This pattern is consistent with trends in medical imaging AI, where convolutional neural networks have largely replaced traditional machine learning methods across diagnostic tasks (30, 90–92). Transition from handcrafted radiomics to deep learning architectures Our analysis shows a clear temporal shift in modeling strategies. Studies from 2019 to 2022 mainly used handcrafted radiomics features or Hounsfield unit measurements with conventional machine learning classifiers. Since 2023, deep learning architectures such as ResNet, DenseNet, and U-Net have become dominant. These approaches enable end-to-end learning, reduce reliance on manual feature engineering, limit interobserver variability in ROI delineation, and capture image patterns beyond predefined radiomic features (93, 94). In 2025, studies began to fine-tune large pretrained models, including OpenCLIP, DINOv2, and CheXagent, indicating a move toward foundation models. This trend aligns with advances in natural language processing and computer vision, where pretrained models adapted to specific tasks often outperform models trained from scratch (95, 96). For osteoporosis screening, such models may learn robust representations of vertebral structure and trabecular texture, improving performance in small datasets and enhancing generalizability across imaging protocols. This progression also introduces limitations. Deep learning models, especially those using whole image inputs without predefined ROIs, often lack interpretability. Although several studies reported excellent performance using whole-image approaches (AUCs 0.82–0.99), these models often provide limited insight into the image regions driving predictions. Without explicit vertebral localization, predictions may be influenced by non-causal image characteristics, reducing model transparency. In contrast, radiomics-based models, despite slightly lower performance, retain greater feature-level interpretability and facilitate biological and clinical interpretation. Predictive value of clinical features Several high-quality studies showed that imaging-only models achieved performance comparable to or better than models that combine imaging and clinical variables, suggesting that deep learning models may implicitly infer age, sex, and certain laboratory derangements directly from images. However, such findings do not imply that clinical features are redundant. A recent study demonstrated that a model relying exclusively on 35 routine clinical variables—without any imaging input—achieved 83.94% accuracy for osteoporosis screening across multiple hospitals (87). This illustrates that clinical features carry substantial independent discriminative power. The modest gains observed in earlier multimodal models may therefore reflect ceiling effects when imaging already captures the relevant physiological information, rather than an inherent limitation of clinical data. Future work should systematically explore both standalone clinical models and fusion architectures, and define the specific contexts—such as suboptimal image quality, limited scanner access, or atypical disease presentations—where clinical variables offer the greatest added value. Performance heterogeneity and task complexity Model performance varied widely across studies, with AUC for osteoporosis prediction ranging from 0.767 to 1.000. This variability reflects both differences in task complexity and methodological factors. Principal component analysis identified participant number and multicenter design as key drivers, with smaller single-center studies consistently reporting higher performance. This pattern indicates overfitting and optimism bias in studies with limited sample size and data diversity. Reports of near-perfect performance require caution. Eight studies achieved AUC ≥ 0.98, including four with an AUC of 1.000. Perfect classification in medical diagnosis is exceptionally rare, given inherent biological variability and measurement error in reference standards (97). These findings likely reflect overfitting, spectrum effects, or limited case heterogeneity rather than true diagnostic perfection. Performance also differed by task. Bone density classification, such as osteoporosis diagnosis based on a T score threshold, provides a relatively consistent reference. In contrast, fracture risk prediction is inherently probabilistic and influenced by multiple factors beyond bone mineral density, including biomechanics and tissue properties (98, 99). This complexity is reflected in lower and more variable performance, with AUC values as low as 0.63. Notably, the only prospectively evaluated model for fracture prediction reported the lowest performance, suggesting that prospective studies yield more realistic estimates than retrospective analyses. Validation gap as a barrier to clinical translation Although PROBAST considers absence of external validation a high risk of bias, only 20 studies (36%) included any external validation, and independent multi-institutional evaluation was rare. The single international validation study illustrates both feasibility and scarcity. This gap directly affects reported performance. Models tested on data from the same institution, scanner, and population tend to overestimate performance compared with evaluations in new settings. Beyond the absence of external validation itself, scanner- and protocol-related heterogeneity constitutes an additional threat to model generalizability. Attenuation values can vary substantially across tube voltages and scanner models, undermining the portability of threshold-based or attenuation-based screening tools. A recent landmark study directly tackled this issue by developing a statistical correction method to normalize trabecular attenuation across 43 scanner models and six tube voltages, using a dataset of over 538,000 CT examinations (100). The automated ROI placement achieved greater than 99% agreement with manual review, and the derived normative values and diagnostic thresholds demonstrated consistent performance across diverse populations, establishing a reproducible framework for multi-scanner opportunistic screening. This study illustrates that rigorous, large-scale validation with explicit scanner calibration is not only possible but essential for clinical translation. Nonetheless, such comprehensive efforts remain the exception. Model development requires substantial resources for data curation, annotation, and training, while data sharing is constrained by regulatory and technical barriers. Without the level of validation and standardization, the clinical utility of opportunistic screening tools will remain uncertain. Future research should therefore prioritize multi-center, multi-scanner designs with built-in calibration strategies to bridge the gap between promising research and real-world deployment. Limitations of this review This review has several limitations. First, although a pre-specified hierarchical extraction strategy was applied to select a single AUC from each of the 57 included studies, this approach may still have introduced a degree of optimistic bias. Specifically, the final last-resort criterion, “selecting the highest reported AUC,” was required in only 8 studies (14.0%). In most studies, the AUC explicitly reported in the abstract was extracted, while external validation results were consistently prioritized over internal validation whenever both were available. Among the 8 studies requiring the last-resort criterion, Study ID 7 reported multiple external validation cohorts without identifying a preferred validation result; Studies 9 and 13 performed external validation but did not clearly distinguish the external validation cohort in the reported results; Study 55 compared clinical, radiomics, and combined clinical–radiomics models without specifying a primary model; Study 33 reported two independent test cohorts without indicating which should be considered the primary validation cohort; and only 3 studies required direct selection of the numerically highest reported AUC without additional contextual guidance. Therefore, although selecting the highest AUC may have introduced some optimistic bias, this occurred in only a small minority of included studies and largely reflected inconsistent reporting rather than arbitrary selection. Because this review aimed to provide a descriptive overview rather than a quantitative meta-analysis of diagnostic performance, we believe that the potential influence of this approach on our overall conclusions is likely to be modest. Nevertheless, the extracted AUC values should be interpreted cautiously, as they may represent upper-bound estimates of model performance rather than the performance typically expected in routine clinical practice. Second, study-level heterogeneity precluded formal meta-analysis. Therefore, we used PCA to illustrate the sources of heterogeneity across studies. Third, inconsistent reporting across studies, particularly for model calibration, clinical utility, and exclusion criteria, limited the assessment of model quality. Fourth, the focus on model development rather than validation studies means that reported performance likely reflects upper bounds rather than real-world effectiveness. Fifth, although our search strategy (see Supplementary Table 2) combined multiple keywords (including “opportunistic,” “incidental,” “unsuspected,” and “asymptomatic”) to balance sensitivity and specificity, some functionally opportunistic studies may still have been missed. For instance, a relevant MRI study (86) that came to our attention during the review process was not captured by the initial search, indicating that certain articles using routine imaging for bone density assessment without these specific terms in their titles or abstracts may have been inadvertently excluded. Sixth, only English-language publications were included, which may have introduced both language and publication bias. Given the rapid expansion of AI research in medical imaging across non-English-speaking countries, relevant studies published in local-language journals may have been missed. This restriction may reduce the global representativeness of the available evidence and underrepresent variations in imaging protocols, healthcare settings, and patient populations. Furthermore, studies reporting favorable model performance are more likely to be published in international English-language journals, potentially leading to an overly optimistic assessment of the current diagnostic performance and clinical readiness of AI-based opportunistic osteoporosis screening. This residual gap should be considered when interpreting the scope of our findings. Conclusion This review highlights the rapid growth of AI-based opportunistic screening for osteoporosis, marked by a shift toward deep learning and more advanced modeling approaches. However, progress in model development has not been matched by advances in validation. The dominance of retrospective, single-center studies and the lack of external validation remain key barriers to clinical translation. Although current evidence supports technical feasibility, the next phase requires multi-institutional collaboration, prospective evaluation, and standardized reporting. These steps are essential to translate current models into reliable clinical tools for osteoporosis detection and fracture prevention.